Skip to content

Physical AI agents: OWASP coverage gap for robotic/actuator systems #787

@pshkv

Description

@pshkv

The toolkit covers the software agent security surface well. Opening this to discuss a gap that the current seven packages don't address: physical AI agents — LLM-driven systems that actuate in the real world via ROS 2, MAVLink (drones), industrial controllers, or embedded hardware.

Why physical agents require additional governance primitives

Software agents that go wrong can be rolled back. Physical agents that go wrong cause irreversible harm — a drone that arms without authorization, a robot arm that exceeds safe velocity near a human, a welding robot that actuates without human sign-off.

The OWASP Agentic Top 10 framework covers these conceptually but the enforcement mechanism is different:

OWASP Category Software agent Physical agent
Tool Misuse (OAT-05) Block the API call Block the hardware command before it reaches the actuator
Cascading Failures (OAT-08) Rate limit HTTP calls Cap collective kinetic energy Σ(½mv²) across a drone fleet
Privilege Escalation (OAT-02) Scope check on API token Velocity/force constraint in signed capability token

Specific gaps in the current toolkit

  1. No physical constraint model — There's no equivalent of maxVelocityMps, maxForceNewtons, or geofence in the policy schema. Physical agents need these as first-class constraints, not just action allow/deny.

  2. No environment-adaptive constraints — Static policy rules can't adapt to real-time sensor state. An obstacle at 0.4m should automatically cap velocity to 0.1 m/s regardless of what the token says. This requires a plugin interface that runs on every action, not a static policy file.

  3. No tier-based irreversibility modelARM + MISSION_START on a drone is categorically different from a sensor read. The governance layer needs to know that some actions are physically irreversible and require explicit human sign-off before execution (not audit after).

  4. No swarm collective constraints — Per-agent governance doesn't capture emergent collective risk. A fleet of individually-authorized drones can collectively exceed safe kinetic energy limits.

What we've built as a reference

SINT Protocol is an open-source physical AI authorization layer that covers exactly this surface:

  • Tier model: T0_observe → T1_prepare → T2_act → T3_commit mapped to reversibility + physical risk
  • MAVLink bridge: 15 tier rules (ARM → T3, TAKEOFF → T2, camera → T0)
  • ROS 2 bridge: topic-level capability token enforcement
  • DynamicEnvelopePlugin: min(token.maxVelocityMps, obstacle_distance × reaction_factor) per action
  • SwarmCoordinator: collective KE ceiling, minimum inter-agent distance, concurrent actor limit
  • Evidence ledger: SHA-256 hash chain with TEE attestation (SGX/TrustZone/SEV)

950 tests across 18 packages.

Proposal

Happy to collaborate on:

  1. Adding a physical agent profile to the OWASP coverage matrix
  2. A PhysicalConstraintPolicy schema extension for agent-os
  3. A reference integration between SINT's bridge-mavlink / bridge-ros2 and the AGT policy engine

Is physical AI in scope for the toolkit's roadmap? Would the team be open to an integration PR?

Metadata

Metadata

Assignees

Labels

enhancementNew feature or request

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions